NvMayMay/weedblaster-vision-yolov8s

WeedBlaster Vision Model — YOLOv8s on CropAndWeed

Detection backbone for a laser-powered autonomous weed-killing robot. Identifies 8 individual crop species and a weed superclass in real time to enable precision laser ablation without chemicals.

Overview

This is a YOLOv8-s (small) model fine-tuned via transfer learning on the CropAndWeed dataset (Steininger et al., WACV 2023). The model uses the CropsOrWeed9 label mapping: 8 individual crop classes plus a single weed superclass, optimized for the real-world task of distinguishing "what to protect" from "what to destroy."

Why this matters: Herbicides account for ~$30B in global ag-chem spend annually and are linked to soil degradation, water contamination, and biodiversity loss. Precision laser weeding eliminates chemical use entirely by targeting individual weed plants at the stem — but only if the vision system can reliably distinguish crops from weeds in real time, across variable field conditions.

Model Performance

Version: v1.9 · Architecture: YOLOv8-s · Input Resolution: 1280×1280 px · Augmentation: Enabled

Class	Precision
Overall	0.960
Bean	0.988
Sunflower	0.985
Pumpkin	0.980
Pea	0.975
Maize	0.973
Sugar Beet	0.973
Soy	0.950
Potato	0.920
Weed (superclass)	0.850

The model achieves 96% overall precision, meaning fewer than 4% of detections are false positives — critical for a system that fires a laser at whatever it classifies as a weed.

Training Curves

Dataset

CropAndWeed Dataset — Steininger et al., WACV 2023.

• 8,034 annotated images from 929 recording sessions across Austrian agricultural sites and experimental plots
• ~112k annotated plant instances across 74 species (16 crops, 58 weeds)
• Bounding boxes, semantic masks, and stem positions
• High variability in lighting, soil type, moisture, and growth stage
• Top-down capture at ~1.1m height with 50mm focal length — representative of robot-mounted camera perspectives

We use the CropsOrWeed9 variant, which maps the 74 original classes into 8 crop species (Maize, Sugar Beet, Soy, Sunflower, Potato, Pea, Bean, Pumpkin) and a single Weed superclass aggregating all 58 weed species.

Training Configuration

Parameter	Value
Base model	yolov8s.pt (COCO pretrained)
Epochs	100
Image size	1280×1280 px
Batch size	8
Optimizer	AdamW
Learning rate	0.001 → cosine anneal to 0.00001 (lrf=0.01)
Weight decay	0.0005
Warmup	3 epochs
Precision	AMP (FP16)
Hardware	NVIDIA Quadro RTX 6000 (24 GB)

Data Augmentation

Augmentation was tuned for agricultural top-down imagery. Rotations and perspective transforms are disabled to preserve plant orientation; mosaic and copy-paste are enabled to improve small-object detection for weeds.

Augmentation	Value	Rationale
HSV Hue	±0.015	Simulate lighting variation
HSV Saturation	±0.7	Soil color / moisture differences
HSV Value	±0.4	Shadow and exposure changes
Scale	±0.5	Growth-stage variation
Translate	±0.1	Camera positioning jitter
Horizontal Flip	0.5	Safe for top-down crop views
Vertical Flip	0.0	Crops have vertical orientation
Rotation	0.0°	Preserves row structure
Mosaic	1.0	Excellent for small weed instances
Mixup	0.1	Helps with overlapping boundaries
Copy-Paste	0.1	Synthetic weed density variation

Quick Start

from ultralytics import YOLO

# Load model
model = YOLO("best.pt")

# Run inference
results = model.predict("field_image.jpg", imgsz=1280, conf=0.25)

# Access detections
for result in results:
    boxes = result.boxes
    for box in boxes:
        cls = int(box.cls[0])
        conf = float(box.conf[0])
        class_names = ['Maize', 'Sugar Beet', 'Soy', 'Sunflower',
                       'Potato', 'Pea', 'Bean', 'Pumpkin', 'Weed']
        print(f"{class_names[cls]}: {conf:.3f}")

Intended Use

This model is the perception component of an autonomous laser weeding system. The operational pipeline is:

1. Detect — this model identifies all crop and weed instances in the camera frame
2. Classify — crops are protected; weeds are targeted
3. Localize — bounding box centers (or future stem-point regression) guide laser aim
4. Ablate — a focused laser destroys the weed at the stem, no chemicals required

Limitations

• Trained on Austrian agricultural data; generalization to other geographies, soil types, and crop varieties requires validation
• Weed superclass does not differentiate between weed species — sufficient for "destroy all weeds" but not for selective herbicide application
• Performance on tiny instances (<16² px) is limited, consistent with dataset design
• Not validated for safety-critical deployment without additional testing and redundancy

Citation

If you use this model, please cite the underlying dataset:

@inproceedings{steininger2023cropandweed,
  title={The CropAndWeed Dataset: A Multi-Modal Learning Approach for Efficient Crop and Weed Manipulation},
  author={Steininger, Daniel and Trondl, Andreas and Croonen, Gerardus and Simon, Julia and Widhalm, Verena},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages={3729--3738},
  year={2023}
}

License

This model is released under AGPL-3.0. The CropAndWeed dataset is available for academic use — see the dataset repository for terms.